v0.9.3

• /api/specs/ingest now reliably auto-schedules the Planner (TFactory #347). Ingested specs were landing at status=pending with planner_scheduled: false and the warning "planner module not importable" — but only in the long-lived server process. The request-time lazy import from agents.planner import schedule_planner (inside create_spec_ingest_workspace) intermittently raised ImportError mid-request, while the very same import resolved cleanly in every fresh process and at app startup. Fix: routes/specs.py now imports agents.planner once at module load (app boot), pinning it into sys.modules so the downstream lazy import is a fast cache hit instead of a fragile fresh import. This unblocks the AIFactory→TFactory verify leg of the PARR loop (the contract already arrives; now the test pipeline actually runs).
• Planner scheduling failures are no longer swallowed. Both schedule_planner call sites in task_control.py now catch Exception (not just ImportError), log the traceback, and surface a planner scheduling failed: … warning — so a future regression shows up in logs and the ingest response instead of a silent pending.

v0.9.2

• /api/specs/ingest now accepts the signed Task Contract and uses it as the authoritative test profile (#71 Phase 3, with AIFactory v3.6.14). The handoff was contract-less ({project_id, spec_id, spec_text}), so the Planner inferred lanes/frameworks and PFactory's declared tfactory block (lanes/frameworks/ac_to_code_map) was discarded. SpecIngestRequest now carries an optional contract; create_spec_ingest_workspace persists it to context/task_contract.json (where read_task_contract() looks first) when it has the RFC-0002 markers, so parse_tfactory_profile drives lane/framework selection. No contract → inference, unchanged (backward compatible).

v0.9.1 — AIFactory→TFactory spec ingest fix

Resolves /api/specs/ingest projects by id-or-name from the web-server store, fixing the AIFactory→TFactory handoff (#517 / #330). Image deployed via the sha-based CD (deploy.yml); this tag is the version marker. CHANGELOG: ## 0.9.1.

v0.9.0

First slice of the enterprise 90-day plan (.agent-os/product/enterprise-90day-plan.md),
plus the per-user token + Ollama Cloud groundwork.

• PR-native quality gate (WS1, #310–#312). The Triager can publish a GitHub
  commit status (TFactory / tests) that passes/fails a PR against a configurable
  policy (agents/quality_gate.py + tools/pr_status.py), opt-in via the
  .tfactory.yml quality_gate block + TFACTORY_PR_STATUS=1. See guides/pr-gate.md.
• Run TFactory without AIFactory — generic spec ingestion (WS2, #313–#315).
  create_spec_ingest_workspace + the task_create_from_spec MCP tool +
  POST /api/specs/ingest (with a typed ingestSpec client) turn a raw
  markdown/Gherkin/EARS spec into a native test-gen task — no branch required.
• GitHub issue → native TFactory test task (#326/#327). The portal's
  issue→task flow now creates a native test-generation task via the ingest
  endpoint, not an inherited AIFactory coding task.
• Tenant hygiene — projects toward org-scoped DB (WS3, #316–#319). A
  projects.json→DB migration into the owner's Personal org, a ProjectStore
  abstraction (JSON default / org-scoped DB behind APP_PROJECTS_BACKEND), a
  request→org resolver, and the live projects route routed through the store seam.
• Java coverage wired into the Evaluator (WS4, #320). Format-aware coverage
  parsing dispatches JaCoCo for the Java lane (Cobertura stays the default).
• Per-user acw_ API tokens on the REST surface (#305). A user-minted key
  with the api:full scope authenticates on /api/* (for the handover skill +
  CLI), with expiry enforcement + last_used_at tracking.
• Ollama Cloud documented + verified (#306). Reachable via the
  openai-compatible:<model> path + OPENAI_COMPATIBLE_* env; connectivity
  checker python -m providers.ollama_cloud_check. See guides/ollama-cloud.md.

v0.8.2

• Agent bash no longer breaks under the OS sandbox on k3d (AIFactory #363). Mirrors AIFactory v3.6.9: bwrap can't mount /proc on k3d (the node is a container), so the SDK's bash sandbox broke every agent command. Gated sandbox.enabled behind AIFACTORY_BASH_SANDBOX (default on); set false on the cluster — bash works, isolation via the K8s pod boundary + command allowlist until gVisor lands.

v0.8.1

• Agent command sandbox now actually engages (#321). The Chainguard runtime
  image omitted bubblewrap (bwrap) and socat, so the Claude Agent SDK logged
  "Sandbox disabled: … bubblewrap (bwrap) not installed" and ran agent bash
  commands with no filesystem/network enforcement — unacceptable for a tool
  that runs generated tests sandboxed. Both are now installed in the runtime
  apk layer. Verified on the cluster: the node allows unprivileged user
  namespaces, so bwrap creates a real sandbox (not just silencing the warning).

v0.8.0

TFactory stops being a standalone tool and becomes a verified node in the
Factory line — it picks up governed test
targets from PFactory, emits a shared completion event CFactory can watch, and
registers itself in the Backstage software catalog. Part of the Factory
PARR-spine epic (Plan · Act · Review · Report).

• CI-parity verification signal — real imports, not mocks (#302). A sixth
  Evaluator signal, inspired by the Hermes agent, that guards against "green
  that lies" — tests passing locally (against mocks or a developer-shaped env)
  yet failing in CI. The pytest lane now grades under a CI-matching environment
  (ambient credentials blanked, TZ=UTC, hash seed pinned, locale normalised)
  on top of the existing --network=none --read-only sandbox, and a static
  "real-imports" check flags any suite passing only by mocking out the subject
  module under test (signals_summary.ci_parity: mocked-subject → demote
  accept→flag). Surfaced in the verdict + triage report; documented in
  guides/testing-model.md.
• PFactory governed-target pickup (#193–#197). TFactory now consumes test
  work that PFactory has planned and
  governed, rather than only ad-hoc handovers:
  • Recognise + enqueue (#195). Detect governed test targets and enqueue
    them through the existing task pipeline.
  • Tag taxonomy (#194). Create the shared label set PFactory and TFactory
    agree on for routing governed work.
  • pfactory:meta oracle (#196). Parse the pfactory:meta block as the
    test oracle — the governed acceptance contract the generated tests must
    satisfy.
  • Generate · run · report back (#197). Produce tests for a picked-up
    target, execute them, and report the result back up the spine.
• RFC-0001 normalized completion event (#198 / #211 / #214 / #224). The
  Triager's terminal-status completion event now conforms to
  RFC-0001
  — the canonical cross-service correlation-key + completion-event schema, so
  one shape flows across AIFactory · PFactory · TFactory and CFactory watches a
  single contract:
  • Shared correlation_key (GitHub issue number, with a synthetic
    tf-<spec_id> fallback so it is never null); legacy int correlation_id
    retained as a back-compat alias.
  • Default integration port moved 3102 → 3103 (#198); MCP-proxy tests
    realigned.
  • Triager emits an RFC-0001 usage block on completion events (#224).
  • See docs/completion-event-envelope.md.
• Backstage software-catalog onboarding (#215 / #216 / #223). TFactory is now
  importable into Backstage: catalog-info.yaml + TechDocs (#215/#216),
  completion-event TechDocs aligned to RFC-0001, and enriched catalog
  annotations + an AI-assistant skill descriptor (#223).
• Reach more systems under test. Multi-step / SSO test-target login flows
  (#107) and toHaveScreenshot visual baselines wired to the portal-managed
  store (#109); live Kubernetes port-forward dispatch fix (#108).
• Hardening. Patched the binutils CVE (CVE-2026-6846) to clear the P0 Trivy
  gate; web-server deps installed into the backend venv so the full suite imports
  cleanly (#219); core pipeline agents decomposed per a Clean Code review;
  pre-existing ruff lint debt cleared. Backend suite now at 2803 tests.

v0.7.0

Hardens the Factory PARR spine's completion-event delivery (epic #284, TFactory
side complete), fixes the OIDC SSO login loop, and rounds out the SaaS connector
support.

✨ New Features

• Completion outbox + retrying relay (#281). At-least-once delivery of
  RFC-0001 completion events — the Triager durably enqueues the envelope before
  delivery; a relay drains it with exponential backoff + dead-lettering and
  replays across restarts. Opt-in via TFACTORY_COMPLETION_OUTBOX /
  APP_COMPLETION_RELAY_ENABLED.
• Additive envelope upgrade (#282). Per-event id (idempotency key),
  CloudEvents-core (specversion/source/type/time), and W3C traceparent
  ride alongside the legacy fields. Published schema + CI validation.
• Typed handback contract + bounded retry (#283). Versioned triage-report
  contract (contracts/handback-triage-contract.v1.schema.json), assertion
  pinning (manifest hash + additive-only diff-gate), and a terminal needs_human
  completion event when the correction-cycle cap is reached.
• SaaS connector visual lane (#173). Opt-in visual: true browser lane on
  ConnectorTarget/HttpTarget for visual-inspection runs, with ServiceNow
  browser guidance (iframe#gsft_main + stable selectors).
• SAP OData connector (#111). sap-odata.py.tmpl api-lane check (OData v2/v4,
  bearer/OAuth or Basic) — all four connector platforms now have library checks.

🐛 Bug Fixes

• OIDC/Keycloak SSO login loop (#286). Honor the access_token cookie in the
  auth middleware + get_current_user, and make the SPA checkAuth cookie-aware
  so SSO logins no longer bounce back to /login.
• binutils CVE-2026-6846 (#218). Pinned binutils>2.46-r1; Trivy P0 gate green.

🔧 Other

• Login page auto-versions from package.json at build time.

v0.5.0

Close the loop with AIFactory: when TFactory's tests find problems, hand a
correction back to AIFactory for a fix, then re-test — bounded so it can't
run away. The reverse of /handover-to-tfactory. Epic
#182.

• Hand-back pipeline (#182). New agents/handback/ packages a finished
  run's failures into a QA_FIX_REQUEST.md-shaped correction and (opt-in) sends
  it to AIFactory's QA Fixer:
  • Traceability (#183). The snapshotter records the AIFactory hand-back
    target (aifactory{project_id,spec_id,api_url,task_id} + correction_cycle)
    in context/source.json; api_url defaults to AIFactory's web-server
    (http://localhost:3101, override TFACTORY_AIFACTORY_API_URL).
  • Builder + renderer (#184). request.py + render.py — pure-compute
    selection of failing tests → a deterministic fix-request payload (reuses the
    triage report + the visual correction plan).
  • Sender + Triager hook (#185). send.py writes
    findings/handback_request.{md,json} always and POSTs only on
    dry_run=False AND confirm; the Triager's terminal-status hook prepares
    (default ON) and sends on opt-in (TFACTORY_HANDBACK_SEND=1), mirroring
    TFACTORY_TRIAGER_GIT_WRITE.
  • Operator skill + CLI (#186). /handback-to-aifactory (+ AIFactory
    companion) previews then sends via the AIFactory MCP tool
    task_apply_correction, or python -m agents.handback <spec_dir> --send.
  • Bounded closed loop (#187). loop.py + /tfactory-fixloop drive a
    test→fix→re-test cycle that stops at passed, or stuck (cap
    TFACTORY_HANDBACK_MAX_CYCLES default 2, or no progress).
  • AIFactory receiver (sister repo, AIFactory#317): POST /api/tasks/{task_id}/apply-correction + MCP task_apply_correction write
    QA_FIX_REQUEST.md onto the original spec and run the existing QA Fixer.
  • Dry-run-first + opt-in throughout (no automatic pushes). See
    guides/aifactory-handback.md and
    docs/plans/2026-06-03-aifactory-tfactory-handback-design.md.

v0.4.0

A new Visual Inspection Run feature (all no-tenant phases shipped) plus the
first-class SaaS connector target. Epic
#170.

• Visual Inspection Run (#170). Record a generated Playwright browser run
  (trace + video + step-labeled verification and error screenshots) and package
  it into automated-test/<YYYY-MM-DD-HHMMSS>/ with a deterministic human report,
  an LLM correction plan (injectable seam + deterministic fallback), a GitHub
  issue export (dry-run), and meta.json. New agents/visual_inspection/
  (packager · report · correction_plan · issues · store); a
  write_paths_to_branch git helper to commit the folder to the SUT repo
  (dry-run default); a portal Visual Reports page + /api/visual-inspections
  routes; and a /handover-to-tfactory opt-in (visual_inspection {enabled, target, flow} threaded through task_create_and_run). Phases P1/P2/P4/P5
  shipped; P3 (ServiceNow browser/SSO) remains, needing a live tenant.
  See docs/plans/2026-06-03-visual-inspection-run-design.md.
• SaaS connector target (#111). A first-class type: connector target
  (ServiceNow / Salesforce / SAP / MuleSoft) reusing the http + credential-vault
  auth, plus a platform registry mapping each platform → API style · library/
  check template · guidance. See guides/saas-connectors.md.
• storageState login-once scaffolding (#107). Gen-Functional scaffolds
  auth.setup.ts + a requires_auth Playwright config from a ref-auth target's
  selectors, so a browser test logs in once and reuses the session.